The role of AI in patient care - Compass from Harvard Med School

A recent development from Harvard Medical School caught my attention because it shows both the promise of healthcare AI and the questions that come with it. Researchers developed COMPASS, an AI model designed to predict which cancer patients are more likely to respond to immune checkpoint inhibitors, using patterns in tumor gene activity. In retrospective testing across 16 clinical cohorts, the model outperformed the best existing approach by 8.5%, and importantly, it was designed to provide a rationale for its predictions rather than simply producing a black-box answer.

That is an exciting direction for precision medicine, when the treatment decision involves drugs that may work remarkably well for some patients but not for others. If a system like this eventually moves from research into clinical practice, what level of evidence should be required before its recommendation begins influencing treatment? How should clinicians interpret the model's rationale, and how much confidence should they place in it when an individual patient doesn't resemble the populations on which it was evaluated?

The researchers themselves make an important distinction: COMPASS still needs validation in prospective clinical trials before it could become a clinical decision aid. That distinction is important for healthcare organizations.

So the big question is: “What evidence do we need before we are prepared to act on what it tells us?”

Source: Harvard Medical School, AI Tool Improves Prediction of Who Will Respond to Cancer Immunotherapy Drugs, reporting research published in Nature Medicine, July 2026.
Read the Harvard Medical School article